Adaptive Graph Auto-Encoder for General Data Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Graph Theory

Background:

• Graph neural networks (GNNs) excel with graph-structured data.
• General clustering tasks often lack pre-existing graph structures, limiting GNN applicability.
• Graph construction is critical for GNN-based clustering performance.

Purpose of the Study:

• To extend GNNs to general clustering tasks by adaptively constructing graphs.
• To propose AdaGAE, a graph auto-encoder for general data clustering.
• To address graph degeneration issues in adaptive graph construction.

Main Methods:

• Developed AdaGAE, a graph auto-encoder utilizing a generative graph perspective.
• Implemented an adaptive graph construction process to exploit high-level data information and non-Euclidean structures.
• Designed a novel mechanism to prevent performance collapse during graph updates.

Main Results:

• AdaGAE demonstrates robust and stable performance across diverse datasets.
• The proposed adaptive graph construction effectively utilizes data information.
• The novel mechanism successfully avoids performance degeneration.
• AdaGAE is insensitive to parameter initialization and requires no pretraining.

Conclusions:

• AdaGAE effectively extends graph auto-encoders to general clustering tasks.
• The adaptive graph generation and collapse-prevention mechanism are key innovations.
• AdaGAE offers a stable, efficient, and versatile solution for clustering various data types.